Artificial Intelligence (AI) is, broadly speaking, the idea of mechanised intelligence; something that can reason, learn, or make decisions without an actual brain. The concept itself is far older than computers. Ancient Egyptian, Greek, and Chinese myths all imagined automata endowed with some form of intelligence [1]. Modern fiction hasn't let go of the idea either. From C-3PO and HAL 9000 to Ultron and Optimus Prime, artificial intelligence has become one of science fiction's favourite obsessions. But over the last few years, AI has escaped fiction entirely and forced its way into everyday life. It's helping doctors analyse scans [2] and researchers discover drugs [3]. It's ordering our groceries [4], choosing our entertainment [5], and picking our partners [6]. It's flooding the internet with content [7] and apparently preventing social-media from showing you posts in chronological order [8]. It's powering deepfakes [9], mass surveillance [10], and autonomous weapons systems [11]. Yet despite its ubiquity, nobody can agree on what AI actually is. Part of the problem is that the definition keeps changing. Not long ago, calculators and chess engines were considered artificial intelligence [12]. Today they're just tools. Until a few years ago, AI often meant "things computers can't do yet". Now we're being told that AI has finally arrived. So what exactly has arrived? And is it intelligent?

What is Intelligence?

If we're going to decide whether something is artificially intelligent, it probably helps to figure out what intelligence actually is first. The problem is that nobody really knows. Is intelligence reasoning? Pattern recognition? Learning? Creativity? Is it all of these? Or do these things emerge from intelligence? Philosophers, neuroscientists, mathematicians, computer scientists, and quiz show hosts have been arguing about intelligence for decades and we're still no closer to an answer. Humans don't make this especially easy to pin down either. I managed to write a PhD thesis, but also got caught on a fisherman's hook while diving off a pier because I thought a colourful lure looked interesting. Things get even trickier when questions of self-awareness and understanding get thrown into the mix. Is it enough to solve a problem without understanding the how or why behind the solution? While we may all have an intuitive idea of what intelligence means, modern AI research tends to avoid those bigger questions directly. Instead, it focuses on outcomes. Can a machine learn patterns, adapt to unfamiliar situations, make useful predictions, or produce convincing responses?

Types of AI

As if defining intelligence wasn't difficult enough, there are multiple types of artificial intelligence too. In fact, most of what we've called AI up to now has been extremely specialised.

Rule-based AI

Early AI systems were mostly giant collections of rules. If X happens, do Y. If an email contains certain phrases, mark it as spam. If a patient has particular symptoms, suggest a diagnosis. If the player walks into range, attack. These systems can appear intelligent within their own context, but fall apart the moment reality hits. They don't learn or adapt. They simply follow instructions very quickly.

Narrow AI

This is the AI we actually have today. Narrow AI is designed for one task and one task only. A chess engine can beat world champions but can't make a sandwich. Spotify can recommend music but can't feel the emotion it evokes. Modern machine learning improved things by allowing systems to learn patterns from enormous amounts of data instead of relying entirely on hardcoded rules. Rather than teaching a computer every possible cat shape, you simply show it ten million photos of cats and let it figure things out itself.

AGI

Artificial General Intelligence (AGI) is the hypothetical idea of a "human-level and beyond" AI. This is the kind of AI science fiction warns us about. Unlike today's systems, an AGI wouldn't just perform one task well. It would be able to transfer knowledge between tasks, reason in unfamiliar situations, learn new skills on the fly, and apply what it knows across very different problems. If a system like that ever became more capable than the brightest humans, it could, in theory, lead the process of improving itself. That recursive improvement is often referred to as the "singularity"; a point where technological progress grows exponentially and becomes very difficult to predict.

Current AI systems are not AGI. Despite headlines, marketing, and CEO boasts, ChatGPT cannot suddenly decide to become a builder or overthrow humanity.

Large Language Models

Large Language Models (LLMs) are the reason for the renewed interest in AI. Several big companies are using them to power their products; OpenAI has ChatGPT, Anthropic has Claude, Google has Gemini, etc.. LLMs let you type naturally and they respond in kind. It feels less like using software and more like talking to someone. But underneath all the hype, LLMs are fundamentally prediction engines.

How LLMs Work

Imagine trying to predict the next letter in an English sentence based on only a small amount of context. If you see a "Q", you'd likely guess "U". If you see "TH", "E" is a strong candidate. As you add more context, the game becomes easier. If you've already seen "PREDICTIO", you can be fairly confident what comes next. This is the idea behind Markov models; systems that predict the next state based on the current one, often with a limited memory of the past. Human language behaves in a similar way, except you often need a bigger context than just a few previous characters or words. If I say "Apple and...", you might guess "cinnamon". But if I say "The two tech giants Apple and...", your prediction is more likely to be "Google" or "Microsoft". At a very basic level, LLMs work by doing exactly this: repeatedly predicting the next word in a sequence. Where older approaches, like Markov models, only looked at a small window of previous words, LLMs can use vast amounts of context; entire paragraphs, their tone and structure, and relationships between ideas spread far apart in the text. They are trained on enormous datasets by repeatedly learning to fill in missing words. And when I say repeatedly, they can do it billions upon billions of times. Over time, this produces a model that's extremely good at statistical patterns in language. Good enough to explain quantum physics, write code, argue about films online, and occasionally invent facts with absolute confidence.

This is, of course, a wild simplification, as there's levels of verification tooling that goes into a finished LLM response.

Is it Intelligent?

LLMs don't truly understand anything in the human sense. They don't experience the world, hold beliefs, possess desires, or attach meaning to words. At their core, they are systems trained to predict statistically likely sequences of text. We don't say something just because of the words that preceded. Or do we?

Humans and animals also learn from patterns, memory, and experience. A bird can learn that making certain sounds results in food. A dog might do the same with gestures. Do they understand the meaning of their words and actions as we claim to? I would argue, to a certain degree, yes [13]. An LLM does something very different. It can generate convincing descriptions of emotion, hunger, or pain, without ever experiencing any of them. We can relate to each other's realities, but not to that of the LLM.

This brings us back to the problem of defining intelligence. If intelligence simply means producing useful responses, adapting to context, and solving problems, then modern LLMs are clearly intelligent systems. But if intelligence requires understanding, self-awareness, intention, or subjective experience, then prediction and validation alone surely are not enough. A system can imitate reasoning extremely well without necessarily reasoning in the way humans do.

The Turing Test famously proposed that if a machine could convincingly imitate human conversation, we should treat it as intelligent. Modern LLMs can often do exactly that, but their success may say as much about the limitations of the test as the capabilities of the machines themselves. Sounding human is not the same thing as understanding.

For now, Large Language Models are probably best understood not as artificial minds, but as extraordinarily sophisticated systems for modelling and generating language. That is still an astonishing technological achievement. But it is very different from a machine that thinks, understands, or experiences the world in the way humans do.